WALL-WM

WALL-WM is an open-source world action model for general-purpose embodied AI, developed by X Square Robot (自变量机器人, Shenzhen) and released on 29 May 2026. It belongs to the WALL series and is described in the paper “WALL-WM: Carving World Action Modeling at the Event Joints” (arXiv:2606.01955).

Its key idea is a shift from chunk-centric learning over fixed time windows to event-grounded vision-language-action pretraining, where the atomic unit of learning is a semantically coherent action event (e.g. grasping, lifting). This lets the model learn task objectives rather than memorizing pixel sequences, and better reconciles language semantics, continuous visual dynamics and control-level actions.

The model couples video prediction with action prediction at event boundaries. It offers two inference modes: an event mode (variable-length) and a unified mode with Staircase Decoding. It uses a Muon-optimizer-based pretraining infrastructure and a three-layer architecture.

WALL-WM is released under the Apache-2.0 license together with code and tooling (GitHub wall-x, Hugging Face), permitting commercial and non-commercial use, self-hosting and fine-tuning. It complements the WALL series with world-modeling and action-consequence prediction for planning and robot policy learning.